A total of 134 samples were identified with the following methods: appearance analysis; microscopic identification; thin-layer chromatography (TLC) identification; moisture content; and ash content. These detection items were conducted as described in the BS section of the Chinese Pharmacopoeia (2020) and the BOS sections of the Sichuan Provincial Standards for Processing Chinese Herbal Pieces (2015), the Gansu Provincial Standards for Processing Chinese Herbal Pieces (2009), and the Gansu Provincial Standards for Chinese Medicinal Materials (2009).

GC-IMS was performed with a FlavourSpec^® flavor analyzer (GAS, Germany) equipped with analytical software including Laboratory Analytical Viewer (LAV), Reporter, Gallery Plot, and GC × IMS Library Search. For each sample, 1.0 g of crude powder was accurately weighed and placed in a 20-mL headspace bottle. The powder was incubated at 80°C for 20 min before injection into the flavor analyzer. Each sample was injected twice in parallel. The headspace injection conditions were as follows: incubation temperature, 80°C; incubation time, 20 min; incubation speed, 500 rpm; injection needle temperature, 85°C; injection volume, 500 μL. For GC, an FS-SE-54-CB-1 chromatographic column of 15 m in length was used; the column temperature was 60°C. The carrier gas flow volume was as follows: 0–2 min, 2 mL·min⁻¹; 2–20 min, 2–100 mL·min⁻¹; and 20–30 min, 100 mL·min⁻¹. For IMS conditions, the carrier and drift gas were N₂, the temperature was 45°C, and the drift gas flow rate was 150 mL·min⁻¹. The NIST and IMS databases included with the instrument software were used to identify the volatile organic compounds present in the samples, then the abundances of the volatile organic compounds were analyzed with the Reporter and Gallery Plot functions.

Olfactory information was collected using 10 types of metal oxide sensors (W1C, W5S, W3C, W6S, W5C, W1S, W1W, W2S, W2W, and W3S) in a PEN3 portable electronic nose (AIRSENSE, Germany). For each of the 134 samples, there were three replicates of 2 g of powder each. Based on data from our pre-experimental results, the samples were tested after incubating at room temperature in a covered container for 30 min. The sampling time was 80 s, the cleaning time was 80 s, the sensor zeroing time was 5 s, the sample preparation time was 5 s, and the air intake flow rate was 400 mL·min⁻¹. The resulting olfactory data were used to construct the olfactory information matrix X1 (134 × 10).

The IRIS VA400 electronic eye was used to collect visual data from the samples. Samples of an area of about 10 × 10 cm² were randomly selected and placed on A4 white paper. Top lighting conditions were selected based on pre-experimental results. A 24-color plate was used for color correction. The electronic eye used a 5-mm aperture and the upper and lower backlights were used simultaneously to eliminate the background. Three separate images were taken of each sample with the position of the slices changed between images. The resulting visual data, obtained with 85 sensors, were used to construct the visual information matrix X2 (134 × 85).

Taste information was collected from each of the 134 samples using the TS-5000Z Insent electronic tongue (Ensoul Technology LTD. Electronic tongue sensors include Sourness, Bitterness, Astringency, Aftertaste-B, Aftertaste-A, Umami, Richness and Saltiness.). For each sample, 2 g was weighed out and crushed in an electric homogenizer for 15 s. The resulting powder was placed in a 100-mL beaker with an appropriate volume of purified water. After stirring, samples were incubated at room temperature without perturbation for 5 min. The samples were subsequently filtered, sterilized and poured into a special cup to be tested by the electronic tongue. The electronic tongue sensor was cleaned in a cleaning solution for 90 s, in a reference solution for 120 s, and in a different reference solution for 120 s. The sensor started to collect sample information after the response value stabilized at 0 for 30 s. The acquisition time of the beforetaste value of each sample was 30 s, the sensors were then cleaned for 3 s in the two reference solutions. Finally, the sensors were inserted into the new reference solution to collect data for 30 s and the aftertaste value was exported. This cycle was repeated four times; data from the first cycle were removed and the average value was calculated from the last three cycles. The resulting taste data, obtained from eight sensors, were used to construct the taste information matrix X3 (134 × 8).

We first analyzed 134 dried tuber samples as set forth in the Pharmacopoeia and with microscopic, TLC, moisture content, and ash content analysis. Based on these analyses combined, 45 batches of B. striata decoction pieces met the standards for classification as BS on the Chinese Pharmacopoeia (Part I) and did not meet the standards for classification as BOS based on the Sichuan Provincial Standards for Processing Chinese Herbal Pieces, the Gansu Provincial Standards for Processing Chinese Herbal Pieces, or the Gansu Provincial Standards for Chinese Medicinal Materials. Similarly, 30 samples each of GE and PO decoction pieces did not meet the standards for classification as BS; they were positively identified as authentic GE and PO, respectively. However, many of the 29 batches of BOS not only met the standards for classification as BOS, but also met the standards for classification as BS. Specifically, the traits of BOS samples 106–120 were essentially identical to those of BS. However, samples 121–134 were highly wooded and had a significantly different texture than BS, thus meeting local standards for classification as BOS. There was extremely high similarity in TLC results between the 29 BOS batches and the control medicinal samples of BS, making it difficult to distinguish between the two sample types. Microscopic characteristics were consistent between the 29 BOS batches and the BS samples, displaying the characteristic “epidermal cells with wavy curved walls, calcium oxalate needle crystal bundles, catheters, fiber bundles and gelatinized starch granules” described in Pharmacopoeia. However, the microscopic background was more turbid in BOS than in BS samples. Overall, the moisture contents of 134 samples ranged from 6.10% to 12.75% and the ash contents ranged from 1.22% to 4.39%, consistent with expectations based on the Pharmacopoeia.

We next analyzed the levels of volatile organic compounds in each sample with GC-IMS. Sample 34 (BS), sample 49 (GE), sample 79 (PO), and sample 108 (BOS) were randomly selected for comparison of the two-dimension vertical view plot of BS with the vertical view plots of the other sample types (Figure 2). In Figure 2, the ordinate represents the GC retention time and the abscissa represents the normalized ion draft time. The red vertical line at abscissa = 1.0 is the normalized reactive ion peak (RIP). Each point on the right side of the RIP peak represents one type of volatile organic compound. Concentrations are indicated with color depth. The results clearly showed significant differences in both the types and abundances of volatile organic compounds between BS and other three types of decoction pieces.

We first compared the fingerprints of BS and GE (Figure 3A). In regions A and B, levels of volatile compounds were higher in BS. Specifically, BS had an increased abundance of butyraldehyde, 3-methylbutyric acid, 2,3-pentanedione, 4-methylthiazole, 2-butanone, 2-heptenal, 2-hexenal (monomer), 2-hexenal (polymer), and several unknown compounds compared to GE. Region C shows volatile compounds that were more abundant in GE than in BS, namely, ethyl propionate, 3-pentanone, acetone, acetic acid, 2,3-butanedione, 2-methyl-1-pentanol, furfural (monomer and dimer), phenylacetaldehyde, 2-methylbutanal, 5-methylfurfural, and 2-furanmethanol (monomer and dimer).

Comparison of fingerprints between BS and PO (Figure 3B) showed numerous volatile compounds that were more abundant in BS (regions A and B): 2,3-pentanedione, 4-hydroxy-2,5-dimethyl-3 (2H) furanone (dimer), methylheptenone, furfuryl alcohol (monomer), trans-2-nonenal, 2-heptanone, 4-methyl-2-pentanone (dimer), and some unknown compounds. However, compounds including 1-octen-3-ol, furfuryl mercaptan, cis-2-penten-1-ol, isovaleric acid, 3-octanone, 5-methylfuranaldehyde, camphene, and ethyl acetate were more abundant in PO (region C). Finally, the fingerprints of BS and BOS were compared (Figure 3C). Overall, the volatile organic compounds were similar in both identity and abundance between the two types of decoction pieces, although there were subtle differences. Regions A–F show compounds (including 2,3-pentanedione, furfuryl mercaptan, 3-hydroxy-2-butanone, 3-methyl-3-buten-1-ol, 4-methylthiazole, methylheptenone, 4-hydroxy-2,5-dimethyl-3 (2H) furanone [dimer], and 1-penten-3-one) that were more abundant in BS than in BOS. Regions G–I show compounds that were significantly more abundant in BOS than in BS, including benzaldehyde (dimer), eucalyptol, ethenyl acetate, 5-methyl-2-furanmethanol (monomer and dimer), 2,3-butanedione, isobutanol, and 4-methyl-2-pentanone (monomer). Thus, the integrated results of the Pharmacopoeia, local standards, and GC-IMS positively identified samples 1–45 as BS, samples 46–75 as GE, samples 76–105 as PO, and samples 106–134 as BOS, consistent with the known identities of the samples.

The PLS algorithm was used to convert the electronic nose data into linear LVs. Six LVs were then used to establish a PLS-DA model with good performance. These first six LVs explained 99% of the total variance among the samples. For the first two LVs, there were some similarities between BS and counterfeit samples (Figure 6A). In the calibration set, the BS samples S34 and S44 were misclassified, and S76 (PO) was misclassified as BS (Table 1). In the validation set, S57 (GE) was misclassified as BS. Misclassification of the samples may have been due to differences in volatile components compared to other samples resulting from changes in temperature and humidity during transportation and storage. Overall, the model accuracy was 97% in the calibration set and 97.06% in the validation set. The model sensitivity (Se) and specificity (Sp) were 0.94 and 0.98, respectively. This indicated that electronic nose data could be used to accurately distinguish between BS and similar decoction pieces.

We next optimized the kernel parameter and cost values, which were initially 0.05 and 0.1, respectively (Supplementary Figure S2). The error rate in cross-validation changed along with the kernel parameters and cost values, and we selected the combination that produced the lowest error rate and highest accuracy (kernel parameter = 1.13, cost value = 10). The model was run again with the optimized parameters. In the calibration set, S28, S34, and S44 (BS) were misclassified, whereas S64 (GE), S83, S86, and S101 (PO), and S121 (BOS) were misclassified as BS; in the validation set, S30, S81, S82, and S128 were misclassified. Thus, the accuracy was 92% and 88.24% in the calibration and validation sets, respectively. The model Se and Sp were 0.91 and 0.92, respectively. Visualization of the SVM classification hyperplane and decision boundaries (Supplementary Figure S3A) indicated that the performance of the model built with electronic nose data required improvement.

The optimized BP-NN model parameters were as follows: two hidden layers; 10 neurons per layer; learning rate = 0.01; momentum term = 0.5; 500 iterations. After 500 iterations of training, the model error rate was 0 (Figure 7A). In the calibration set, the BS samples S28 and S44 were misclassified; in the validation set, S82 was unassigned. Notably, these three samples were also misclassified or unclassified in the PLS-DA and SVM models, presumably due to differences in sample quality. The accuracy was 98% and 97.06% in the calibration and validation sets, respectively, and the Se and Sp were 0.94 and 1.0, respectively. Thus, the BP-NN model built with electronic nose data could be used to accurately distinguish between BS and related decoction pieces.

In leave-one-out cross-validation, the PLS-DA model performed best when using the first three LVs, which together explained 63% of the sample variance. The two types of samples (BS and others) had good aggregation in the first two LVs (Figure 6B). S84, S91, and S5 were misclassified. Analysis of the original samples suggested that they may have been misclassified because they had slightly different surface color characteristics compared with other samples. The accuracy was 98% and 97.06% in the calibration and validation sets, respectively, and the Se and Sp were 1.0 and 0.97, respectively.

The SVM model had the lowest error rates when the kernel parameter was 1.6, 2.26, 3.2, or 4.53 and the corresponding cost value was 1, 1, 1, or 0.1, respectively; these parameter combinations produced accuracies of 96%, 99%, 99%, and 96%, respectively. Larger kernel parameter and cost values were associated with smaller numbers of support vectors (Supplementary Figure S4). The role of a support vector is to determine the optimal classification hyperplane. The number of support vectors is affected by both the penalty parameter c and the kernel parameter g (Yang and Gao, 2020). A larger c corresponds to fewer support vectors, higher model accuracy, a smaller occupied network space, and faster prediction speeds. The number of support vectors was smallest (51) when the kernel parameter was 3.2 and the cost value was 1. This model was therefore selected for further analysis. Using these parameters, S91 and S5 were misclassified in the calibration and validation sets, respectively; Ac was 99% and 97.06% in the calibration and validation sets, respectively; Se and Sp were 1.0 and 0.98, respectively.

Because there were 85 color values detected by the electronic eye and the input variables were large, the number of hidden-layer neurons in the BP-NN model needed to be optimized for these data. Too many hidden-layer neurons will weaken model generalizability and increase network operation costs. After optimization, the parameters were as follows: three hidden layers; 10 neurons per layer; learning rate = 0.01; momentum term = 0.5; 500 iterations. When the number of training iterations was 500, the model error rate was 0 (Figure 7B). However, S5 was still misclassified in the validation set. This sample comprised irregular thick slices with sections of white, gray-white, and brown; thus, the sample characteristics were different from the irregular thin slices and white color of the 44 other BS samples. In the validation set, S116 was misclassified. Ac was 99% and 97.06% in the calibration and validation sets, respectively and the model Se and Sp were 1.0 and 0.98, respectively.

The PLS-DA model performed best when it was built with four LVs that explained 83% of the sample variance (Figure 6C). S3 and S60 were misclassified in the calibration and validation sets, respectively. BS and non-BS samples showed some similarities in the variance of the first two LVs. The Se and Sp of the model were 0.97 and 1, respectively, and the classification performance was good.

In cross-validation, the SVM model error rate was smallest when the kernel parameter was 0.4, 0.57, 0.8, 1.6, or 3.2, and the corresponding cost value was 1, 1, 1, 10, or 1, respectively. Model accuracy was highest (90% and 94.12% for the calibration and training sets, respectively) at kernel parameter = 0.57 and cost value = 1. Using this model, there were seven and three misclassified BS and GE samples, respectively, in the calibration set. In the validation set, S10 and S67 were misclassified. Se and Sp were 1.0 and 0.98, respectively. The classification result showed that when the SVM classifier performed pattern recognition on the electronic tongue data, the relationship between the recognition mechanism and the differences in electronic tongue data demonstrated a need for further optimization.

The optimal BP-NN model contained two hidden layers and five neurons per layer; it had a learning rate of 0.01, a momentum term of 0.3, and was trained for 500 iterations. In the calibration set, the accuracy was 98%, S15 was unassigned, and S53 was misclassified. S48 was misclassified in the validation set (Ac: 97.06%). After 500 iterations of training, the error rate was 0 (Figure 7C). Model accuracy increased the most when the momentum term was gradually increased from 0.2 to 0.5 while other parameters were unchanged. However, when the momentum term was increased from 0.4 to 0.5, the accuracy decreased due to an increase in the number of unclassified samples. This indicated that a larger momentum term was not well suited for model classification and that parameters in the BP-NN model should be further optimized. The model Se and Sp were 1.0 and 0.98, respectively.

The first 12 LVs, which could explain 89% of the sample variance, were selected to construct a PLS-DA model. In the calibration set, S41 and S119 were misclassified, but the accuracy in the validation set was 100% (Figure 6D). The model Se and Sp were 0.97 and 0.98, respectively, and the classification performance was good. For the SVM model (Supplementary Figure S3D), accuracy was highest (94% and 100% in the calibration and validation sets, respectively) when the kernel parameter was 9 and the cost value was 1. The model Se and Sp were 0.85, 0.98, respectively. The optimized BP-NN model contained two hidden layers and 10 neurons per layer; the learning rate was 0.1, the momentum term was 0.5, and it was trained for 500 iterations, producing an error rate of 0 (Figure 7D). There were five misclassified samples in the calibration set with Ac of 95%, but the validation set accuracy was 100%. The model Se and Sp were 0.91, 0.97, respectively. The high accuracy in the validation set may have been because there were extensive differences in the HPLC data between the four types of decoction pieces; especially when the sample size was small, this could correspond to clear difference between sample types. Thus, for the HPLC data, the accuracies of both the linear and nonlinear classifiers were 100% on the validation set.

The PC scores of the electronic sensor and HPLC data were fused and models were built as described above. The accuracy of the resulting PLS-DA model was 100% for the calibration set, which was superior to the performance of the PLS-DA models constructed with single-source data. In the validation set, only S5 was misclassified (Ac: 97.06%). Se and Sp were 1.0, 1.0, respectively. The two sample types could be completely divided into two categories based on the first two LVs (Figure 8A); the aggregation trend was also superior compared to the models built with single-source data. This indicated that fused PCs (each explaining more than 90% of sample variance in the single-source data) could be used to accurately identify the two sample types while reducing the algorithm running speed. The optimal SVM model had a kernel parameter of 6.4 and cost value of 1; this produced a classification prediction probability of 1 for both sample types (Supplementary Figure S5), which was better than the corresponding SVM model for each single-data source. Model accuracy was also 100% when the kernel parameter was 2.26 or 9 and the cost value was 1, but the number of support vectors was larger in the former case and the classification prediction probability was lower in the latter case. Thus, a kernel parameter/cost value combination of 6.4/1 was selected. The classification hyperplane and support vectors of SVM model was shown in Figure 8C. In the validation set, S5 was misclassified (Ac: 97.06%). Se and Sp were 1.00, 1.00, respectively. The optimal BP-NN model built with the fusion data contained two hidden layers and 10 neurons per layer with a learning rate of 0.1, a momentum term of 0.4. It was trained for 500 iterations, producing a final error rate of 0 (Figure 9A). The area under the curve (AUC) of the receiver operating characteristic (ROC) was 1 (Figure 9B). In the calibration set, S120 was misclassified with Ac of 99.00%; S5 remained misclassified in the validation set with Ac of 97.06%. The model Se and Sp were 1.00, 0.98, respectively.

After fusing the LVs from each data source, all three model types could discriminate between samples better than the corresponding models trained on single-source data and performed comparably to or better than the models built with PC fusion data. The PLS-DA model showed 100% accuracy in both the calibration and validation sets, and Se and Sp were both 1. The two sample types could be completely separated on the first two LVs, and samples clustered together better than in the results of the models built with PC fusion data (Figure 8B). For SVM, the optimal parameter combination was kernel parameter = 9 and cost value = 1 (Figure 8E); this model had 33 support vectors (Figure 8D) and accuracy was 100% on both the calibration and validation sets. Se and Sp were 1.00, 1.00, respectively. The optimized BP-NN model had two hidden layers, 10 neurons per layer, a learning rate of 0.1, a momentum term of 0.3, and 500 iterations (final error rate = 0) (Figure 9C). Using this parameter combination, only S76 was misclassified. The AUC of ROC was 1, and the Se and Sp values at varying prediction probability thresholds showed that the Sp value increases when it is close to 0 and the Se value decreases when it is close to 1 and the area surrounded by the red and blue lines is also large, which indicated that the model had good classification performance (Figure 9D). The accuracy of the final model was 99%, with only one sample misclassified in the validation set and the model Se and Sp were 1.00, 0.98, respectively; this was superior to the classification results based on electronic nose or electronic tongue data alone.

Six LVs which can explain 99% of the sample variance were selected for construction of an optimal PLS-DA species identification model. The four sample types showed similarities in the variance represented by the first two latent variables (Figure 10A). S28 and S34 were misclassified in the calibration set. Because PLS-DA can divide samples only into predefined categories, this model was unable to identify specific samples; one, one, and four BS, GE, and PO samples, respectively, were unassigned. The accuracy was 92% in the calibration set, and Se and Sp were 0.94 and 1, respectively. There were four unassigned samples in the validation set. This indicated a need for further improvement of this model.

The model identification criteria were strict in the BP-NN model because the samples were divided into four categories. This necessitated an appropriate increase in the number of hidden layers and the number of neurons in each layer. The final model contained three hidden layers with 10 neurons per layer and had a learning rate of 0.1, a momentum term of 0.3, and was trained for 500 iterations. In the calibration set, S28 was misclassified and S80 and S83 were unclassified and Ac was 97.00%. The model Se and Sp were 0.97, 1.00, respectively. In the validation set, two samples were misclassified and the overall accuracy was 94.12%.

The first 13 LVs, which explained 95% of the sample variance, were selected to establish a PLS-DA model. The four sample types could be clearly distinguished in the two-dimensional plot, with GE and PO clustering relatively close to one another and BS and BOS samples also clustering near one another (Figure 10B). In the calibration set, S52, S63, and S91 were misclassified and seven samples were unclassified (Ac: 90%). In the validation set, S5, S11, and S48 were unassigned and S67 and S93 were misclassified (Ac: 85.29%). Se and Sp were 1.00, 0.98, respectively.

For the BP-NN model, the optimal parameter combination was two hidden layers, 15 neurons per layer, a learning rate of 0.1, a momentum term of 0.3, and 500 training iterations. In the calibration set, the accuracy was 98%; S23 was unclassified and S69 was misclassified. The model Se and Sp were 1.00, 1.00, respectively. The AUC-ROC was 1, indicating good parameter optimization (Supplementary Figure S6). Increasing the number of hidden layers from two to three while keeping the other parameters unchanged led to a sudden increase in the number of unclassified samples. This suggested that the multi-classification ability of the model could not be improved overall by only increasing the number of hidden layers.

To some extent, taste responses reflected the chemical composition of each sample. The electronic tongue data could therefore theoretically characterize the differences between the four types of decoction pieces better than the other sensory data types. Seven LVs accounting for a full 100% of the sample variance were used to build a PLS-DA model. Compared to the other two sensory data types, the number of unassigned samples decreased to one GE and two PO samples. There were no misclassified samples. Ac was 97.00% and 100% in the calibration and validation sets, respectively and Se and Sp were 1.0 and 1.0, respectively. Clustering based on just the first two LVs also showed good sample separation (Figure 10C); PO samples clustered far away from the other three sample types. BS, GE, and some BOS samples clustered close together, indicating that their tastes were similar. The clear distinction of the latter three sample types from PO was likely because PO is a member of the family Liliaceae, whereas the other three belong to the family Orchidaceae.

The optimal BP-NN model contained two hidden layers with five neurons per layer; the learning rate was 0.1, the momentum term was 0.4, and there were 500 iterations. In the calibration set, S29 and S53 were misclassified and S24, S79, S114 were unclassified (Ac: 95%). In the validation set, the accuracy was 97.06% and there was only one unassigned sample. The model Se and Sp were 0.97, 0.98, respectively. The error rate reached 0 after 500 iterations and the AUC-ROC was 1 (Supplementary Figure S7). This indicated that the selected parameters were appropriate for distinction between the four sample types.

A PLS-DA model was constructed from the first nine LVs, which explained 84% of the sample variance. There were four misclassified and four unclassified samples in the calibration set (Ac: 92%); in the validation set, the accuracy was 100%. Se and Sp were 0.94, 0.98, respectively. In the two-dimensional plot, the BS and BOS samples overlapped, whereas GE and PO samples could be clearly classified into two separate categories, distinct from the BS and BOS samples (Figure 10D). This was consistent with the high similarity between BS and BOS samples in the HPLC fingerprint data. The optimal BP-NN model built on these data had two hidden layers, 10 neurons per layer, a learning rate of 0.1, a momentum term of 0.5, and 500 iterations. There were nine misclassified samples in the calibration set, but the accuracy was 100% in the validation set. The model Se and Sp were 0.94, 0.97, respectively.

The PCs from the electronic sensor data were next fused with the PCs from HPLC data (Table 2). PLS-DA and BN-NN models built with the fused data had improved classification abilities compared to the corresponding models built with single-source data. The PLS-DA model was constructed with the first five LVs. The accuracy in the calibration set was 100%; in the validation set, S5 was still misclassified (Ac: 97.06%). Se and Sp were 1.00, 1.00, respectively. In the variance represented by the first two LVs, the four sample types could be clearly distinguished, although BS and GE were close together (Figure 11A). The optimal BP-NN model had three hidden layers, 10 neurons per layer, a learning rate of 0.1, a momentum term of 0.4, and 500 iterations. After data fusion, there were two unassigned samples in the calibration set (Ac: 97.00%); Se and Sp were 1.0 and 1.0, respectively; the error rate decreased consistently as the training iteration number increased (Figure 11C). The AUC-ROC was 1 for the four sample types (Supplementary Figure S8), indicating that appropriate parameters had been selected.

The LVs were then fused for the electronic sensor and HPLC data. The classification results of a PLS-DA model built with the fused LV data were the same as the results of the model built with the fused PC data: no unclassified or misclassified samples in the calibration set; 100% accuracy; and only S5 unclassified in the validation set. Se and Sp were 1.00, 1.00, respectively. The four sample types were fully separated based on the first two LVs (Figure 11B), and the model performed better than it did with the fused PCs. However, the classification performance of a BP-NN model built with the fused LV data was not significantly improved compared to models built with single-source data. The optimal model had two hidden layers, 15 neurons per layer, a learning rate of 0.1, a momentum term of 0.4, and was trained for 500 iterations. In the calibration set, this model misclassified one sample and two others were unassigned (Ac: 97%); there was one unassigned and one misclassified sample in the validation set (Ac: 94.12%). Se and Sp were 1.00, 1.00, respectively. After 500 training iterations, the error rate was 0 (Figure 11D).